lacommedia/coin-boxing-source-classifier

Coin boxing source-item classifier v2.8

This repository publishes the reproducible classifier specification used to create the Coin boxing curriculum proposals. It includes the exact deterministic rules, prompt, label mapping, generation settings, checkpoint checksum, and completion audit.

Artifact boundary

This is a directed inference pipeline, not a fine-tuned model. No trained weights are included. The referenced local checkpoint was Ternary-Bonsai-2-9B-PTQ1_0.gguf with SHA-256 d11d367066813107d56f3d0a67362f934b80407a39d86f5bc216fc30929c831c. This repository does not redistribute that third-party checkpoint or claim to license it.

The pipeline applies narrow deterministic source rules first. It sends unresolved items and their immutable parent section to an OpenAI-compatible text-generation endpoint. The model returns one letter. An application-side evidence gate can abstain or require review. Every published proposal remains unreviewed and ineligible for runtime promotion.

Results

The completed job processed all 1,773 items with zero execution failures: 893 deterministic results, 699 model calls, and 181 cache materializations. MLflow experiment 40 contained 1,773 finished traces with no missing or duplicate traces. These are pipeline integrity results. They are not human-label accuracy metrics. No validated accuracy, precision, recall, or calibration score exists yet.

Actual and estimated dollar costs are unknown. The run recorded 117,376 prompt tokens and 699 completion tokens for 118,075 total tokens, but local GB10 inference did not have a validated dollar allocation.

Use

classifier.py provides the rules, model request, and letter decoding. Use the companion dataset for error analysis and human review. Do not treat its model proposals as ground truth or use them for technique claims.

Companion dataset: lacommedia/coin-boxing-workout-source-items